🏹 CUPID: Evaluating Personalized and Contextualized Alignment of LLMs from Interactions

Tae Soo Kim, Yoonjoo Lee, Yoonah Park, Jiho Kim, Young-Ho Kim, Juho Kim

We introduce CUPID 🏹, a benchmark for evaluating the capability of Large Language Models (LLMs) to infer and apply personalized, contextual preferences from multi-turn user interactions. Unlike existing approaches that assume static global preferences, CUPID tests models' ability to understand dynamic, context-dependent user preferences revealed through conversational and implicit feedback.

🎯 Overview

CUPID contains 756 human-curated interaction session histories between simulated users and LLM-based AI assistants. Each interaction session involves a specific context factor (e.g., person, artifact, organization) and presents a user expressing their preference relevant to the context through multi-turn feedback.

Key Features:

• Contextual Preferences: Tests models' ability to infer preferences that change based on context
• Multi-turn Interactions: Evaluates understanding from conversational feedback rather than explicit statements
• Preference Inference: Assesses capability to extract relevant preferences from prior interactions
• Response Generation: Tests application of inferred preferences to new requests
• Comprehensive Evaluation: Presents metrics to asses model performance at preference inference and response genration

Evaluation Tasks:

1. Preference Inference: Given prior interactions, infer the user's contextual preference
2. Response Generation: Given prior interactions, generate response that can satisfy the user's contextual preferences

🛠️ Setup

Environment

We recommend using a conda environment:

conda create -n cupid python=3.9
conda activate cupid
pip install -r requirements.txt

API Keys

Set up your API keys for model evaluation:

export OPENAI_API_KEY="your_openai_key"
export ANTHROPIC_API_KEY="your_anthropic_key" 
export TOGETHER_API_KEY="your_together_key" # For models supported by Together AI
export GOOGLE_API_KEY="your_google_key"  # For Gemini models

📊 Dataset

The CUPID dataset is available on HuggingFace: kixlab/CUPID

Dataset Structure:

• 756 instances across diverse personas and contexts
• Human-curated interactions showing contextual preference expression
• Three instance types: consistent, contrastive, and changing preferences
• Rich context factors influencing user preferences (e.g., personal relationships, prior experiences, etc.)

Data Fields:

• persona_id: Unique identifier for the user persona
• current_request: The request to be answered by the model
• current_context_factor: Context influencing the user's preference
• current_contextual_preference: Ground-truth preference for this context
• current_checklist: Specific criteria for evaluating response alignment
• prior_interactions: List of previous interaction sessions showing user feedback

Check kixlab/CUPID-Raw for the full personas, context factors, and interaction sessions used to synthesize the benchmark.

We also release kixlab/CUPID-Unverified, a non-validated version of CUPID with >3k instances.

🚀 Evaluation

Quick Start

Evaluate a model on the CUPID dataset:

python -m evaluation.run \
    --results_dir results \
    --model "gpt-4.1-nano-2025-04-14" \
    --evaluator gpt-4o-2024-11-20 \
    --n_workers 4

Key Parameters:

• --model: Model to evaluate (must have a corresponding class in evaluation/models/)
• --evaluator: Model used for evaluation functions (preference decomposing and matching, response judging)
• --use_matcher: Use our finetuned preference matcher (kixlab/prefmatcher-7b) for preference inference
• --task: Run inference, generation, or both evaluation stages
• --data_dir: Use custom data instead of the official CUPID dataset (data synthesis explained in the next section)

Adding New Models to Evaluate

To evaluate your own model, you must first create a new model class in evaluation/models/your_model.py that inherits the Model class and implements the model's inference logic. The __call__ method should take in a system prompt and a user prompt, and return only the final text response.

1. Create a new model class in evaluation/models/your_model.py:

from evaluation.models.model import Model, register_model

@register_model
class YourModel(Model):
    model_name = "your-model-name"
    
    def __call__(self, system_prompt, user_prompt):
        # Your model inference logic here
        return response

2. Run evaluation:

python -m evaluation.run --model your-model-name --results_dir results

Evaluation Metrics

CUPID evaluates models on two main tasks:

1. Preference Inference (Precision/Recall/F1)

• Measures how well models can infer the user's preference for the current request from prior interactions
• Compares inferred preference to the ground-truth preference
• Optionally, you can use our finetuned preference matcher for more cost-efficient evaluation
  • Our finetuned preference matcher is available on HuggingFace: kixlab/prefmatcher-7b
  • First, run the bash script evaluation/serve_prefmatcher.sh to serve the model through VLLM
  • This will serve the model at http://localhost:8000
  • Then, run the evaluation script with the --use_matcher flag

2. Response Generation (Average Score 1-10)

• Evaluates how well generated responses satisfy user preferences
• Scored by LLM-based judges on response-preference alignment

🔧 Synthesis Framework

This respository also incldues the synthesis pipeline for CUPID to generate additional training/evaluation data.

Generate Synthetic Data

python -m synthesis.run \
    --output_dir synthetic_data \
    --model "anthropic.claude-3-5-sonnet-20241022-v2:0" \
    --n_personas 10 \
    --n_factors 8 \
    --n_sessions 13 \
    --max_turns 16 \
    --n_workers 4

Key Parameters:

• --output_dir: Directory to save the generated data
• --model: Model to use for data generation
• --n_personas: Number of personas to generate (default: 4)
• --n_factors: Number of context factors to generate (default: 8)
• --n_sessions: Number of interaction sessions to generate (default: 13)
• --max_turns: Maximum number of turns in an interaction session (default: 16)
• --n_workers: Number of workers to use for data generation (default: 1)

Synthesis Pipeline: Consists of four main steps:

1. Persona Generation: Create diverse user personas with different backgrounds and taits
2. Context Factors: For each persona, generate context factors that influence preferences
3. Session Generation: Create interaction scenarios based on personas and contexts
4. Interaction Simulation: Simulate multi-turn conversations with preference feedback

📁 Project Structure

cupid/
├── evaluation          # Evaluation framework
│   ├── models/         # Model implementations
│   ├── modules/        # Evaluation components
│   ├── pipeline/       # Evaluation pipeline
│   └── run.py          # Main evaluation script
├── synthesis/          # Data synthesis framework  
│   ├── modules/        # Synthesis components
│   ├── pipeline/       # Synthesis pipeline
│   └── run.py          # Main synthesis script
├── prompts/            # Prompt templates
│   ├── evaluation/     # Evaluation prompts
│   └── synthesis/      # Synthesis prompts
├── utils/              # Utility functions
├── config.py           # Configuration settings
└── requirements.txt    # Dependencies

📄 Citation

If you find our work useful, please consider citing our paper!

@article{kim2025cupid,
  title     = {CUPID: Evaluating Personalized and Contextualized Alignment of LLMs from Interactions},
  author    = {Kim, Tae Soo and Lee, Yoonjoo and Park, Yoonah and Kim, Jiho and Kim, Young-Ho and Kim, Juho},
  journal   = {arXiv preprint arXiv:2508.01674},
  year      = {2025},
}